[Rebase] Rebase to vllm v0.24.0 (#4709)

tests/model_executor/models/moss_tts/test_per_request_codes.py

@@ -156,14 +156,19 @@ class TestPerRequestRouting:
         assert torch.all(audio[0][1] == 0)  # this step's row 0 appended
         assert torch.all(audio[1] == 1)
 
-    def test_runs_without_spans_backward_compat(self):
-        """Missing request_token_spans must not crash (equal-split fallback)
-        and still yields a batch-aligned list."""
+    def test_runs_with_spans(self):
+        """request_token_spans are required for correct per-request routing;
+        verify that passing them produces a batch-aligned list."""
         talker = _make_bare_delay_talker()
         hidden = _row_indexed_hidden(2)
         info_dicts = [{}, {}]
+        spans = [(0, 1), (1, 2)]
 
-        result = talker.make_omni_output(hidden, runtime_additional_information=info_dicts)
+        result = talker.make_omni_output(
+            hidden,
+            runtime_additional_information=info_dicts,
+            request_token_spans=spans,
+        )
         audio = _audio_payload(result)
         assert isinstance(audio, list)
         assert len(audio) == 2

tests/model_executor/models/qwen3_tts/test_qwen3_tts_talker_preprocess.py

@@ -41,6 +41,7 @@ def _make_minimal_talker(
     def _default_raise(**_kwargs):
         raise AssertionError("build_prompt_embeds was not stubbed in this test")
 
+    model._embedding_dtype = torch.bfloat16
     model._prompt_builder = SimpleNamespace(
         build_prompt_embeds=build_prompt_embeds if build_prompt_embeds is not None else _default_raise,
     )
@@ -89,6 +90,7 @@ def _make_minimal_builder(
     )
     builder._speaker_cache = None
     builder._text_tokenizer = None
+    builder._embedding_dtype = torch.bfloat16
     builder._ref_audio_artifact_cache_max_entries = 256
     builder._ref_audio_artifact_cache = OrderedDict()
     builder._resampler_cache = OrderedDict()